Artificial Intelligence Driven Supply Chain Risk Forecasting and Optimization for Resilient Industrial Production Networks

Abstract

Industrial production networks depend on many suppliers whose availability is uncertain. This article asks whether machine-learning forecasts of supplier disruption, passed into a stochastic procurement optimiser, improve the resilience and cost of a multi-tier production network relative to deterministic planning, rule-based buffering and simpler risk scores. Because supplier-level disruption data linked to external signals are not publicly available, the study builds a synthetic network of 8 regions, 80 suppliers, 10 components, 3 plants and 4 products over 312 weeks, in which disruptions arise from a documented latent hazard driven by financial stress, regional risk, weather, port congestion, utilisation and quality, plus correlated regional events. LightGBM, XGBoost, random forests, logistic regression and a neural network were trained on 22 lagged observable features to forecast disruption one to four weeks ahead, calibrated by Platt scaling and explained with TreeSHAP. The calibrated forecasts generated scenario sets for a two-stage stochastic linear program that allocates weekly orders across suppliers with expediting, shortage and surplus recourse and an optional conditional-value-at-risk term. Six procurement policies were compared in a rolling-horizon simulation over a 104-week test window in eight replications with common random numbers. LightGBM reached a test AUC of 0.91 at one week and 0.84 at two weeks, falling to 0.71 at four weeks; logistic regression was within 0.01 AUC and the difference was not significant. The forecast-informed policy reduced total cost by 3.0% on average relative to deterministic planning (paired t = −3.56, p = 0.009), raised fill rate by 1.5 percentage points and cut stock-out weeks from 43 to 17, but its advantage over a persistence policy that reacts to visible outages was small (0.3% of cost). Risk-averse variants raised fill rate further at 0.4–0.7% higher cost in one replication. All results describe the simulated network. The study contributes a reproducible testbed, a decision-focused evaluation protocol and an implementation framework, and cautions that the value of AI here comes mainly from calibrated probabilities inside the optimiser rather than from forecast accuracy alone.